Intel Benchmarks Show 350W GPU D270 Beating NVIDIA’s RTX PRO 6000 By 5.9×
The headline number is that 5.9× figure, and it comes from a direct comparison of the two GPUs across four different models used for options pricing benchmarks: Monte Carlo simulations (both American and European models), Binomial option pricing, and classical Black-Scholes modeling for a European market, the difference being that European markets don't allow you to exercise the option early, before it expires.
The European Monte Carlo simulation is the source of the big 5.9× figure, while the American Monte Carlo simulation is the slowest of the four at "only" 1.92× faster on Crescent Island. The gap between the Intel GPU and the NVIDIA GPU comes down to two things; for one, Intel's Xe3P has full-rate FP64, meaning that it doesn't take a massive performance hit when running double-precision math the way NVIDIA's GPU does. The other advantage it has is fully 2/3 more onboard memory, with 160GB of LPDDR5X versus the 96GB of GDDR7 on NVIDIA's card.
Meanwhile, for the seismic imaging application, the demonstration is less about the performance and more about the ease of porting the application to GPU D270, with "no host-side code changes." Intel and its partner Viridien say that the Earth data company was able to port its seismic imaging mini-application to the GPU D270 in "less than one month" while maintaining both numerical accuracy and "performance," although it's not clear what kind of performance was actually achieved.

This is obviously a big deal because Intel's card tops out at 350W, while NVIDIA's much bigger (in terms of compute resources) GPU can draw up to 600 watts. Of course, this is the same trick AMD pulled with the Ryzen AI Max processors, claiming superior throughput to a big NVIDIA GPU from a relatively small (and integrated) GPU by virtue of being uniquely suited to the task; in that case, by having more memory, and in this case, both having more memory and being designed for a task that the NVIDIA GPU wasn't, really.
NVIDIA's GPU was designed for AI, naturally, but don't worry—Intel's GPU can do that too. Intel claims that the D270 can beat the RTX PRO 6000 in terms of latency for agentic AI decision-making by some 29%. It also says that the D270 offers superior AI compute efficiency versus NVIDIA's GPU, by offering 12% higher LLM throughput at the same 350W power limit (tacitly acknowledging that NVIDIA's part is much faster at full power.)

Intel posits a hybrid compute solution for HPC labs using AI tools that combines SambaNova SN50 Reconfigurable Dataflow Units with GPU D270s and Xeon 6(+) processors to "speed up the entire agentic loop" by intelligently distributing work of different types to the disparate processors in the cluster. Intel brags that this solution is all air-cooled, too, so you don't need expensive liquid-cooling infrastructure.
The firm's showing off the combination of the GPU D270 and SN50 RDUs at the SC26 Supercomputing Conference in Chicago, which runs through the whole week, so if you're headed up that way, make sure to stop by and check it out.

